EP1189203B1 - Homophonewahl in der Spracherkennung - Google Patents

Homophonewahl in der Spracherkennung Download PDF

Info

Publication number
EP1189203B1
EP1189203B1 EP01307935A EP01307935A EP1189203B1 EP 1189203 B1 EP1189203 B1 EP 1189203B1 EP 01307935 A EP01307935 A EP 01307935A EP 01307935 A EP01307935 A EP 01307935A EP 1189203 B1 EP1189203 B1 EP 1189203B1
Authority
EP
European Patent Office
Prior art keywords
words
word
homophone
recognition result
affix
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Expired - Lifetime
Application number
EP01307935A
Other languages
English (en)
French (fr)
Other versions
EP1189203A3 (de
EP1189203A2 (de
Inventor
Stijn Van Even
Olivier Divay
Francoise Renaud
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
L & H Holdings Usa Inc
Original Assignee
L & H Holdings Usa Inc
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by L & H Holdings Usa Inc filed Critical L & H Holdings Usa Inc
Publication of EP1189203A2 publication Critical patent/EP1189203A2/de
Publication of EP1189203A3 publication Critical patent/EP1189203A3/de
Application granted granted Critical
Publication of EP1189203B1 publication Critical patent/EP1189203B1/de
Anticipated expiration legal-status Critical
Expired - Lifetime legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G10—MUSICAL INSTRUMENTS; ACOUSTICS
    • G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
    • G10L15/00—Speech recognition
    • G10L15/22—Procedures used during a speech recognition process, e.g. man-machine dialogue

Definitions

  • This invention relates to computer-implemented speech recognition, and more particularly to selection of homophones in utterances.
  • a speech recognition system analyzes a user's speech to determine what the user said. Most speech recognition systems are frame-based. In a frame-based system, a processor divides a signal descriptive of the speech to be recognized into a series of digital frames, each of which corresponds to a small time increment of the speech.
  • a speech recognition system may be a "discrete” system that recognizes discrete words or phrases but which requires the user to pause briefly between each discrete word or phrase.
  • a speech recognition system may be a "continuous” system that can recognize spoken words or phrases regardless of whether the user pauses between them. Continuous speech recognition systems typically have a higher incidence of recognition errors in comparison to discrete recognition systems due to complexities of recognizing continuous speech.
  • the processor of a continuous speech recognition system analyzes "utterances" of speech.
  • An utterance includes a variable number of frames and corresponds, for example, to a period of speech followed by a pause of at least a predetermined duration.
  • the processor determines what the user said by finding acoustic models that best match the digital frames of an utterance, and by identifying text that corresponds to those acoustic models.
  • An acoustic model may correspond to a word, phrase or command from a vocabulary.
  • An acoustic model also may represent a sound, or phoneme, that corresponds to a portion of a word. Collectively, the constituent phonemes for a word represent the phonetic spelling of the word. Acoustic models also may represent silence and various types of environmental noise.
  • the processor may identify text that corresponds to the best. matching acoustic models by reference to phonetic word models in an active vocabulary of words and phrases.
  • the processor may produce a single recognition candidate for an utterance, or may produce a list of recognition candidates.
  • Incorrect text associated with recognition errors in computer-implemented speech recognition is corrected by performing speech recognition on an utterance to produce a recognition result for the utterance.
  • a correction command is identified in the recognition result for the utterance.
  • the correction command includes a set of one or more words from a prior recognition result, and an affix that may be combined with the set of one or more words to form a homophone of the one or more words.
  • the homophone is generated from the set of one or more words and the affix based on features of one or more words from the prior recognition result that are not included in the set of one or more words.
  • Embodiments may include one or more of the following features.
  • the set of one or more words from the prior recognition result may be replaced with the homophone.
  • Generating the homophone from the set of one or more words and the affix may include combining the set of one or more words and the affix in such a way as to match the features of the one or more words from the prior recognition result with the homophone. Generating the homophone may include presenting a list of possible homophones, and receiving a selection of one homophone from the list.
  • the correction command may include "ecrire ⁇ word> ⁇ ending>," where ⁇ word> represents the set of one or more words from the prior recognition result and ⁇ ending> represents the affix.
  • the set of one or more words may include only one word and the affix is an ending.
  • the homophone may be pronounced exactly the same as the set of one or more words from the prior recognition result.
  • the one or more phonemes of the homophone may be included in the set of one or more words from the prior recognition result.
  • the homophone includes a set of one or more words.
  • the affix may be an ending (that is, a suffix), or the affix may be a prefix.
  • a speech recognition system uses word features and phrases in a recognition candidate to correct homophone errors. As background information for the discussion of correction of homophones, a speech recognition system that does not use a homophone correction command is discussed with reference to Figs. 1-16B.
  • a speech recognition system 100 includes input/output (I/O) devices (for example, microphone 102, mouse 104, keyboard 106, and display 108) and a computer 110 having a processor 112, an I/O unit 114 and a sound card 116.
  • a memory 118 stores data and programs such as an operating system 120, an application program 122 (for example, a word processing program), and speech recognition software 124.
  • a computer suitable for one implementation of the system includes a 700 MHz Pentiumnf processor, 128 MB memory, 12 GB of available hard drive space.
  • the computer may run Windows 95, Windows 98, Windows2000 or Windows NT 4.0 as an operating system or any other suitable operating system.
  • the computer 110 may be used for traditional speech recognition.
  • the microphone 102 receives the user's speech and conveys the speech, in the form of an analog signal, to the sound card 116, which in turn passes the signal through an analog-to-digital (A/D) converter to transform the analog signal into a set of digital samples.
  • A/D analog-to-digital
  • the processor 112 identifies utterances in the user's continuous speech. Utterances are separated from one another by a pause having a sufficiently large, predetermined duration (for example, 160-250 milliseconds). Each utterance may include one or more words of the user's speech.
  • the system also may include an analog recorder port 126 and/or a digital recorder port 128.
  • the analog recorder port 126 is connected to the sound card 116 and is used to transmit speech recorded using an analog or digital hand-held recorder to the sound card.
  • the analog recorder port 126 may be implemented using a line in port.
  • the hand-held recorder is connected to the port using a cable connected between the line in port and a line out or speaker port of the recorder.
  • the analog recorder port 126 may be implemented as a microphone positioned so as to be next to the speaker of the hand-held recorder when the recorder is inserted into the port 126, and also may be implemented using the microphone 102.
  • the analog recorder port 126 may be implemented as a tape player that receives a tape recorded using a hand-held recorder and transmits information recorded on the tape to the sound card 116.
  • the digital recorder port 128 may be implemented to transfer a digital file generated using a hand-held digital recorder 130. This file may be transferred directly into memory 118, or to a storage device such as bard drive 132.
  • the digital recorder port 128 may be implemented as a storage device (for example floppy drive or CD-ROM drive) of the computer 110.
  • the computer 110 runs interface software 140, the speech recognition software 124, a parser 142, and back-end software 144.
  • Dragon NaturallySpeaking Preferred Edition 3.1 available from Dragon Systems, Inc. of Newton, Massachusetts, offers one example of suitable speech recognition software.
  • the interface software 140 provides a user interface for controlling the transfer of data from the digital recorder and the generation of action items for use by the back-end software 144.
  • the user interface may be controlled using input devices such as a mouse or keyboard, or using voice commands processed by the speech recognition software 124.
  • the interface software 140 After transferring data from the recorder, the interface software 140 provides the digital samples for an action item to the speech recognition software 124. If the digital samples have been stored using compression techniques, the interface software 140 decompresses them prior to providing them to the speech recognition software. In general, the speech recognition software analyzes the digital samples to produce a sequence of text, and provides this sequence to the interface software 140. The interface software 140 then transfers the text and the associated time stamp, if any, to the parser 142, which processes the text in conjunction with the time stamp to generate a parsed version of the action item. The parser returns the parsed action item to the interface software, which displays it to the user. After any editing by the user, and with user approval, the interface software then transfers the action item to the appropriate back-end software 144.
  • An example of back-end software with which the system works is personal information management software, such as Microsoft Outlook, which is available from Microsoft Corporation of Redmond, Washington. Other suitable back-end software includes contact management software, time management software, expense reporting applications, electronic mail programs,
  • Fig. 2 illustrates components of the speech recognition software 124.
  • the following description indicates that the components carry out operations to achieve specified results. However, it should be understood that each component actually causes the processor 112 to operate in the specified manner.
  • a front end processing module 200 converts the digital samples 205 from the sound card 116 (or from the digital recorder port 128) into frames of parameters 210 that represent the frequency content of an utterance, Each frame includes 24 parameters and represents a short portion (for example, 10 milliseconds) of the utterance.
  • the front end processing module 200 produces a frame from digital samples according to a procedure 300.
  • the module first produces a frequency domain representation X(f) of the portion of the utterance by performing a Fast Fourier Transform (FFT) on the digital samples (step 305).
  • FFT Fast Fourier Transform
  • the module determines log(X(f)) 2 (step 310).
  • the module may then perform frequency warping (step 315) and a filter bank analysis (step 320) to achieve speaker normalization. See S. Wegmann et al., "Speaker Normalization on Conversational Speech," Proc. 1996 ICASSP, pp. I.339-I.341.
  • the module performs cepstral analysis to produce twelve cepstral parameters (step 325).
  • the module generates the cepstral parameters by performing an inverse cosine transformation on the logarithms of the frequency parameters. Cepstral parameters and cepstral differences (described below) have been found to emphasize information important to speech recognition more effectively than do the frequency parameters.
  • the module After performing channel normalization of the cepstral parameters (step 330), the module produces twelve cepstral differences (that is, the differences between cepstral parameters in successive frames) (step 335) and twelve cepstral second differences (that is, the differences between cepstral differences in successive frames) (step 340).
  • the module performs an IMELDA linear combination transformation to select the twenty-four most useful parameters from the twelve cepstral parameters, the twelve cepstral differences, and the twelve cepstral second differences (step 345).
  • a recognizer 215 receives and processes the frames of an utterance to identify text corresponding to the utterance.
  • the recognizer entertains several hypotheses about the text and associates a score with each hypothesis.
  • the score reflects the probability that a hypothesis corresponds to the user's speech.
  • scores are maintained as negative logarithmic values. Accordingly, a lower score indicates a better match (a high probability) while a higher score indicates a less likely match (a lower probability), with the likelihood of the match decreasing as the score increases.
  • the recognizer provides the best-scoring hypotheses to the control/interface module 220 as a list of recognition candidates, where each recognition candidate corresponds to a hypothesis and has an associated score.
  • Some recognition candidates may correspond to text while other recognition candidates correspond to commands. Commands may include words, phrases, or sentences.
  • the recognizer 215 processes the frames 210 of an utterance in view of one or more constraint grammars 225.
  • a constraint gammar also referred to as a template or restriction rule, may be a limitation on the words that may correspond to an utterance, a limitation on the order or grammatical form of the words, or both.
  • a constraint grammar for menu-manipulation commands may include only entries from the menu (for example, "file”, “edit”) or command words for navigating through the menu (for example, “up”, “down”, "top”, “bottom”). Different constraint grammars may be active at different times.
  • a constraint grammar may be associated with a particular application program 122 and may be activated when the user opens the application program and deactivated when the user doses the application program.
  • the recognizer 215 discards any hypothesis that does not comply with an active constraint grammar.
  • the recognizer 215 may adjust the score of a hypothesis associated with a particular constraint grammar based on characteristics of the constraint grammar.
  • Fig. 4A illustrates an example of a constraint grammar for a "select" command used to select previously recognized text.
  • a constraint grammar may be illustrated as a state diagram 400.
  • the "select" command includes the word “select” followed by one or more previously-recognized words, with the words being in the order in which they were previously recognized.
  • the first state 405 of the constraint grammar indicates that the first word of the select command must be "select”. After the word "select", the constraint grammar permits a transition along a path 410 to a second state 415 that requires the next word in the command to be a previously-recognized word.
  • a path 420 which returns to the second state 415, indicates that the command may include additional previously-recognized words.
  • Fig. 4B illustrates the state diagram 450 of the constraint grammar for the select command when a previously-recognized utterance is "four score and seven". This state diagram could be expanded to include words from additional utterances.
  • the "select" command and techniques for generating its constraint grammar are described further in U.S. Patent No. 5,794,189, titled “CONTINUOUS SPEECH RECOGNITION" and issued August 11, 1998.
  • the constraint grammar also may be expressed in Backus-Naur Form (BNF) or Extended BNF (EBNF).
  • BNF Backus-Naur Form
  • EBNF Extended BNF
  • the grammar for the "Select" command is:
  • this notation indicates that "select” may be followed by any ordered sequence of previously-recognized words.
  • This grammar does not permit optional or alternate words.
  • the grammar may be modified to permit optional words (for example, an optional "and” to permit "four score and seven" or “four score seven") or alternate words or phrases (for example, "four score and seven" or “eighty seven”). Constraint grammars are discussed further in U.S. Patent No. 5,799,279, titled “CONTINUOUS RECOGNITION OF SPEECH AND COMMANDS" and issued August 25, 1998.
  • the large vocabulary dictation grammar identifies words included in the active vocabulary 230, which is the vocabulary of words known to the software.
  • the large vocabulary dictation grammar also indicates the frequency with which words occur.
  • a language model associated with the large vocabulary dictation grammar may be a unigram model that indicates the frequency with which a word occurs independently of context, or a bigram model that indicates the frequency with which a word occurs in the context of a preceding word. For example, a bigram model may indicate that a noun or adjective is more likely to follow the word "the" than is a verb or preposition.
  • Other constraint grammars 225 include an in-line dictation macros grammar for dictation commands, such as "CAP” or “Capitalize” to capitalize a word and "New-Paragraph” to start a new paragraph; the select X Y Z grammar discussed above and used in selecting text; an error correction commands grammar, a dictation editing grammar; an application command and control grammar that may be used to control a particular application program 122; a global command and control grammar that may be used to control the operating system 120 and the speech recognition software 124; a menu and dialog tracking grammar that may be used to manipulate menus; and a keyboard control grammar that permits the use of speech in place of input devices, such as the keyboard 106 or the mouse 104.
  • an in-line dictation macros grammar for dictation commands, such as "CAP” or “Capitalize” to capitalize a word and "New-Paragraph” to start a new paragraph
  • the select X Y Z grammar discussed above and used in selecting text
  • an error correction commands grammar such as "CAP” or
  • the active vocabulary 230 uses a pronunciation model in which each word is represented by a series of phonemes that make up the phonetic spelling of the word.
  • Each phoneme may be represented as a triphone that includes three nodes.
  • a triphone is a context-dependent phoneme. For example, the triphone " abc " represents the phoneme “ b “ in the context of the phonemes " a " and “ c “, with the phoneme “ b “ being preceded by the phoneme " a " and followed by the phoneme " c ".
  • One or more vocabulary files may be associated with each user.
  • the vocabulary files contain all of the words; pronunciations and language model information for the user. Dictation and command grammars may be split between vocabulary files to optimize language model information and memory use, and to keep each single vocabulary file under a size limit, for example, 64,000 words. There also is a set of system vocabularies.
  • acoustic models 235 arc provided for each user of the system. Initially, speaker-independent acoustic models of male or female speech are adapted to a particular user's speech using an enrollment program. The acoustic models may be further adapted as the system is used. The acoustic models are maintained in a file separate from the active vocabulary 230.
  • the acoustic models 235 represent phonemes.
  • the acoustic models 235 represent each triphone node as a mixture of Gaussian probability density functions ("PDFs").
  • PDFs Gaussian probability density functions
  • PDFs probability density function
  • the vectors ⁇ k each include twenty four parameters.
  • the matrices c k are twenty four by twenty four matrices.
  • Each triphone node may be represented as a mixture of up to, for example, sixteen different PDFs.
  • a particular PDF may be used in the representation of multiple triphone nodes.
  • the acoustic models 235 represent each triphone node as a collection of mixture weights w k associated with up to sixteen different PDFs N k and separately represent each PDF N k using a mean vector ⁇ k and a covariance matrix c k .
  • Use of a particular PDF to represent multiple triphone nodes permits the models to include a smaller number of PDFs than would be required if each triphone node included entirely separate PDFs. Since the English language may be roughly represented using 43 different phonemes, there may be up to 79,507 (433) different triphones, which would result in a huge number of PDFs if each triphone node were represented by a separate set ofPDFs.
  • Representing multiple nodes with common PDFs also may remedy or reduce a data sparsity problem that results because some triphones (for example, "tzp" in the English language) rarely occur. These rare triphones may be represented by having closely-related triphones share the same set of PDFs.
  • a large vocabulary dictation grammar may include multiple dictation topics (for example, "medical” or "legal"), each having its own vocabulary file and its own language model.
  • a dictation topic includes a set of words that represents the active vocabulary 230. In a typical example, a topic may include approximately 30,000 words that are considered for normal recognition.
  • a complete dictation vocabulary consists of the active vocabulary 230 plus a backup vocabulary 245.
  • the backup vocabulary may include files that contain user-specific backup vocabulary words and system-wide backup vocabulary words.
  • User-specific backup vocabulary words include words that a user has created while using the speech recognition software. These words are stored in vocabulary files for the user and for the dictation topic, and are available as part of the backup dictionary for the dictation topic regardless of user, and to the user regardless of which dictation topic is being used. For example, if a user is using a medical topic and adds the word "ganglion" to the dictation vocabulary, any other user of the medical topic will have immediate access to the word "ganglion”. In addition, the word will be written into the user-specific backup vocabulary. Then, if the user says “ganglion” while using a legal topic, the word "ganglion" will be available during correction from the backup dictionary.
  • the system-wide backup vocabulary contains all the words known to the system, including words that may currently be in an active vocabulary.
  • the recognizer 215 may operate in parallel with a pre-filtering procedure 240.
  • the recognizer 215 requests from the pre-filtering procedure 240 a list of words that may have been spoken as the first word of the utterance (that is, words that may correspond to the first and subsequent frames of the utterance).
  • the pre-filtering procedure 240 performs a coarse comparison of the sequence of frames with the active vocabulary 230 to identify a subset of the vocabulary for which a more extensive comparison using the recognizer is justified.
  • the pre-filtering procedure 240 uses a lexical tree 500 that is initialized before processing begins.
  • the lexical tree represents the active vocabulary 230 based on the phonetic relafonships between words in the vocabulary.
  • the lexical tree based on the phonetic relationships between words in the vocabulary.
  • the lexical tree includes a root node 505 that represents new words entering the lexical tree. From the root node 505, the tree expands to a group 510 of nodes that correspond to phonemes with which words start. A silence node 512 that represents silence also may be reached from the root node 505.
  • Each node in the group 510 represents a phoneme that appears at the beginning of one or more words.
  • a node 610 corresponds to all words in the vocabulary that start with the phoneme "H".
  • the nodes in the group 510 include representations of the starting phoneme of every word in the vocabulary.
  • leaf nodes 515 that represent the actual words of the vocabulary.
  • leaf node 615 of Fig. 6 corresponds to the word "healing”.
  • An internal node of the tree also may represent a word of the vocabulary.
  • the node 520 might represent a particular vocabulary word in addition to representing the first two phonemes of other vocabulary words.
  • the leaf node 620 of Fig. 6 corresponds to the words “heal” and “heel” while also corresponding to the first three phonemes of the words “heals", “heels” and “healing”.
  • Node 620 also illustrates that, since multiple words may have the same phonetic spelling, a leaf node may correspond to more than one word.
  • leaf nodes may appear at different levels within the lexical tree.
  • Leaf nodes also may correspond to commands.
  • a leaf node may correspond to the word "select” and to the command "SELECT".
  • commands may be associated with particular constraint grammars 225.
  • the pre-filtering procedure begins by retrieving the next frame of parameters for an utterance (step 700). Immediately after initialization, the next frame will be the first frame for the utterance. Thereafter, the next frame will be the frame following the last frame that was processed by the pre-filtering procedure when the pre-filtering procedure was last called.
  • the pre-filtering procedure does not reinitialize the lexical tree between requests for list of words. Accordingly, the state of the lexical tree when a list of words is requested corresponds to the state of the lexical tree after a previous list of words was returned.
  • the pre-filtering procedure finds an active node in the tree with no unprocessed active successors (step 705). Successors of a node also may be referred to as subnodes of the node.
  • the silence node 512 is the only active node.
  • the pre-filtering procedure processes the current node (step 710) according to a node-processing procedure 1100 that is discussed below with reference to Fig. 11.
  • the node-processing procedure determines whether the node should spawn additional active nodes and whether the node should be rendered inactive. If the node is a leaf node, the node-processing procedure also determines whether the word corresponding to the node should be added to a word list for a time associated with the node.
  • the pre-filtering procedure determines whether the node is the highest node in the tree (that is, the root node) (step 715). If the node is not the highest node, then the pre-filtering procedure goes to the next node having no unprocessed active subnodes (step 720) and processes that node (step 710). When searching for the next node to process, the pre-filtering procedure considers inactive nodes having active subnodes or active siblings.
  • the pre-filtering procedure processes the silence node 512 (step 725).
  • the silence node is processed by comparing a frame to a model for silence and adding the resulting score to the minimum of the current score for the silence node and the score for the root node 505.
  • the pre-filtering procedure reseeds the lexical tree (step 730).
  • the pre-filtering procedure reseeds the tree whenever the silence node 512 is active or a word was produced by a leaf node of the lexical tree, regardless of whether the word was added to the list of words.
  • the pre-filtering procedure reseeds the tree by replacing the score for the root node 505 with the minimum of the score for the silence node 512 and the scores for any words produced by leaf nodes of the lexical tree for the current frame. If the silence node is inactive and no leaf node has produced a word, then the pre-filtering procedure replaces the score for the root node 505 with a bad score (that is, a score having a value larger than a pruning threshold).
  • the pre-filtering procedure determines whether more words may be added to the word list for the requested time (step 735). If there are no active nodes in the lexical tree corresponding to speech that started at, before, or slightly after the start time for which the list was requested, and if the last frame to be processed corresponds to a time that is slightly after the start time for which the list was requested, then no more words may be added to the word list. A word produced by the lexical tree is added to the list of words corresponding to the start time of the word and to lists of words corresponding to times that precede and follow the start time of the word. It is for this reason that the pre-filtering procedure waits until there are no active nodes in the tree corresponding to speech that started slightly after the start time for the list of words. If more words may be added, then the pre-filtering procedure retrieves the next frame of parameters (step 700) and repeats the steps discussed above.
  • the pre-filtering procedure returns the word list (step 740) to the recognizer 215. If the word list includes more than a predefined number of words, then the pre-filtering procedure removes words from the list prior to returning the list. The pre-filtering procedure removes the words that are least likely to correspond to the user's speech and removes enough words to reduce the number of words on the list to the predefined number. The procedure also deletes any lists of words for times prior to the requested start time.
  • Each node of the lexical tree 500 represents a sequence of states for a particular phoneme.
  • Fig. 8A illustrates a node 800 that includes a first state 805, a second state 810, and a third state 815.
  • a comparison with a frame of parameters may cause the score in a particular state to remain in the state (through a path 820).
  • a score remains in the state when the score, after being adjusted based on a comparison with a model for the state, is better than a score passed from a preceding state or node, or when no score is passed from a preceding state or node.
  • the comparison also may cause the score to be passed to a subsequent state through a path 825.
  • a score is passed to a subsequent state when the score, after being adjusted based on a comparison with a model for the subsequent state, is better than the score in the subsequent state, or when no score is associated with the subsequent state.
  • the score for the third state 815 may be passed to one or more subsequent nodes through a path 830.
  • the node 512 that corresponds to silence is represented by a single state 840.
  • Each comparison with a frame of parameters may cause a score in the node to remain in the state 840 (through the path 845) and also may cause the score to be passed to the root node 505 through a path 850.
  • the root node 505 is represented by a single state 860. Comparison with a frame causes the score in the node to be passed to one or more subsequent nodes (including the silence node 512) through a path 865.
  • Each state of a node may be represented by four values: a score, a starting time, a leaving penalty, and a staying penalty.
  • the score represents the likelihood that a series of frames has placed the lexical tree in the state (that is, the probability that the series of frames corresponds to the word or portion of a word to which the state corresponds).
  • the scores are maintained as negative logarithmic values.
  • the starting time identifies the hypothesized time at which the user began to speak the word or words represented by the state.
  • the starting time identifies the time at which the score associated with the state entered the lexical tree (that is, the time at which the score was passed from the state 840 along the path 850).
  • the leaving and staying penalties are fixed values associated with the state.
  • the staying penalty is added to any score that stays in the state.
  • the staying penalty is related inversely to the length of the sound represented by the state and to the length of the phoneme represented by the node to which the state belongs.
  • the staying penalty could be proportional to -log(1- 1/d avg ), where d avg is the average duration, in frames, of the sound represented by the state.
  • the staying penalty has a relatively large value when the sound corresponding to the state occurs for only a small amount of time and a relatively small value when the sound corresponding to the state occurs for a large amount of time.
  • the leaving penalty is added to any score that exits the state, and includes a duration component and a language model component.
  • the duration component is related directly to the length of the sound represented by the state and to the length of the phoneme represented by the node to which the state belongs.
  • the duration component of the leaving penalty could be proportional to -log(1/d avg ).
  • the duration component of the leaving penalty has a relatively large value when the sound corresponding to the state occurs for a large amount of time and a relatively small value when the sound corresponding to the state occurs for a small amount of time.
  • the language model components of the leaving penalties for all states in a particular node together represent a language model score for the phoneme associated with that node.
  • the language model score represents the likelihood that a word including the phoneme will occur in speech.
  • the language model score included in the leaving penalties for a node is the increase in the best language model score for the branch of the lexical tree that begins with the node relative to the branch of the lexical tree that begins with the node's parent.
  • the following discussion assumes that there are no leaving or staying penalties associated with the state 840 or the state 860. The same result could be achieved by setting the leaving and staying penalties for states 840 and 860 equal to zero.
  • the following discussion also assumes that the first frame is the first frame that may correspond to speech instead of silence.
  • Fig. 9 provides a simplified example of how scores propagate through the lexical tree.
  • state 840 (which corresponds to silence) has a score of 0 and no other nodes are active.
  • the score of 0 means that there is a one hundred percent probability that the system is starting from silence.
  • the score for the state 840 (S A1 ) is set equal to the acoustic score (A A1 ) resulting from an acoustic match of the first frame with an acoustic model corresponding to the state 840 (that is, the acoustic model for silence).
  • the score for the state 840 (S A1 ) is set equal to the likelihood that the first frame corresponds to silence.
  • Retrieval of the first frame also causes the state 805 to become an active state.
  • the score for the state 805 (S B1 ) is set equal to the acoustic score (A B1 ) resulting from an acoustic match of the first frame with the acoustic model corresponding to the state 805.
  • the score for the state 805 (S B1 ) is set equal to the likelihood that the first frame corresponds to the state 805.
  • the starting time for the state 805 is set equal the time associated with the first frame. This value for the starting time indicates that the score at state 805 represents a word that started at a time corresponding to the first frame. The starting time moves with the score as the score propagates through the lexical tree.
  • each of the scores corresponds to a negative logarithmic probability. Accordingly, adding scores together corresponds to multiplying the probabilities.
  • the score for the state 840 (S A2 ) equals the likelihood that both of the first and second frames correspond to silence.
  • the score for state 805 (S B2 ) corresponds to the more likely of two alternatives: (1) the first frame was silence and the second frame was the sound represented by the state 805 or (2) both of the first and second frames were the sound represented by the state 805.
  • the first alternative corresponds to a transition from state 840 to state 805 along the path 850.
  • the second alternative corresponds to a transition from state 805 back to state 805 along path 820.
  • the starting time corresponding to the first frame that was stored previously for the state 805 is replaced by a value corresponding to the second frame. This value indicates that the score at state 805 represents a word that started with the second frame.
  • the state 810 becomes an active state.
  • leave C and leave D are leaving penalties for, respectively, states 810 and 815.
  • the sum of language model components of leave B , leave C and leave D represents the language model score for the phoneme represented by the node 800.
  • a node may be processed according to a node-processing procedure 1100.
  • the node-processing procedure updates the scores and time values for each state of the node (step 1105).
  • the node-processing procedure updates the scores and time values by generating acoustic scores and using the equations discussed above.
  • the node-processing procedure uses the score for the last state to generate scores for any inactive subnodes of the node. If the generated score for a subnode does not exceed a pruning threshold, then the node-processing procedure activates that subnode and provides the subnode with the generated score.
  • the node-processing procedure determines whether the score of any state of the node exceeds the pruning threshold (step 1110). When a score exceeds the pruning threshold, the likelihood that the word represented by the score was spoken is deemed to be too small to merit further consideration. For this reason, the procedure prunes the lexical tree by deactivating any state having a score that exceeds the pruning threshold (step 1115). If every state of the node is deactivated, then the node-processing procedure also deactivates the node. The node-processing procedure may deactivate a node or state by deleting a record associated with the node or state, or by indicating in the record that the node or state is inactive.
  • the node-processing procedure may activate a node or state by creating a record and associating the record with the node or state, or by indicating in an existing record that the node or state is active.
  • the procedure may use a dynamic pruning threshold that accounts for variations in the average or best score in the lexical tree at any given time.
  • the node-processing procedure determines whether a word is to be added to a list of words (step 1120).
  • a word is added to the list of words when the node being processed corresponds to the last phoneme of a word, a score has been propagated out of the last state of the node, and the score is less than a list threshold.
  • the node-processing procedure adds a language model score to the score.
  • the language model score corresponds to the difference between the language model score for the word and the incremental language model score that is already included in the score.
  • the list threshold has a lower value than the pruning threshold. If the node being processed corresponds to the last phoneme of multiple words having the same phonetic spelling, then all of the words to which the node corresponds are added to the list of words.
  • the node-processing procedure adds the word or words to the list (step 1125).
  • a word is stored in the list of words along with the score propagated out of the last state. If the word is on the list already, then the node-processing procedure stores with the list the better of the score already stored with the list or the score propagated out of the last state.
  • the scores for words in a list of words are returned along with the list of words. The recognizer 215 uses these scores in making the detailed match.
  • the node-processing procedure also adds the word to lists of words for times that precede or follow the starting time to account for possible inaccuracies in the starting time of the word that may result from selecting the better of a score that remains in a state or a score propagated from a prior state. Spreading the word across multiple lists ensures that these inaccuracies will not impinge on the accuracy of the speech recognition system.
  • the node-processing procedure spreads the word across multiple lists based on the length of the word.
  • the node-processing procedure After adding a word to the list of words (step 1125), the node-processing procedure saves the score associated with the word as a reseeding score for use in reseeding the tree (step 1130).
  • Production of a word by the lexical tree means that the current frame may correspond to the last frame of the word (with the probability of such a correspondence being reflected by the score associated with the word). This means that the next frame may correspond to the beginning of a word or to silence resulting from a pause between words.
  • the pre-filtering procedure reseeds the tree (step 730 of Fig. 7) to account for this possibility.
  • the pre-filtering procedure activates the root node 505 and associates the minimum of the reseeding score (S RS ) and the score for the silence node 512 with the root node.
  • the active root node 505 may be used to activate nodes in the group 510 or to activate the silence node 512.
  • Processing of the node is complete after the node-processing procedure saves a score for use in reseeding the tree (step 1130), or if no word is to be added to the list of words (step 1120).
  • the lexical tree pre-filtering procedure is discussed in detail in U.S. Patent No. 5,822,730, titled "LEXICAL TREE PRE-FILTERING IN SPEECH RECOGNITION" and issued October 13, 1998.
  • the recognizer After the pre-filtering procedure responds with the requested list of words, the recognizer initiates a hypothesis for each word from the list and compares acoustic models for the word to the frames of parameters representing the utterance. The recognizer uses the results of these comparisons to generate scores for the hypotheses. Hypotheses having excessive scores are eliminated from further consideration. As noted above, hypotheses that comply with no active constraint grammar also are eliminated.
  • the recognizer determines that a word of a hypothesis has ended, the recognizer requests from the pre-filtering procedure a list of words that may have been spoken just after the ending-time of the word. The recognizer then generates a new hypotheses for each word on the list, where each new hypothesis includes the words of the old hypothesis plus the corresponding new word from the list.
  • the recognizer uses acoustic scores for words of the hypothesis, a language model score that indicates the likelihood that words of the hypothesis are used together, and scores provided for each word of the hypothesis by the pre-filtering procedure.
  • the scores provided by the pre-filtering procedure include components corresponding to a crude acoustic comparison and a language model score indicative of the likelihood that a word is used, independently of context.
  • the recognizer may eliminate any hypothesis that is associated with a constraint grammar (for example, a command hypothesis), but does not comply with the constraint grammar.
  • the recognizer 215 operates according to a procedure 1200.
  • the recognizer 215 initializes the lexical tree 500 as described above (step 1205).
  • the recognizer 215 retrieves a frame of parameters (step 1210) and determines whether there are hypotheses to be considered for the frame (step 1215).
  • the first frame always corresponds to silence so that there are no hypotheses to be considered for the first frame.
  • the recognizer 215 goes to the first hypothesis (step 1220).
  • the recognizer compares the frame to acoustic models 235 for the last word of the hypothesis (step 1225) and, based on the comparison, updates a score associated with the hypothesis (step 1230).
  • the recognizer determines whether the user was likely to have spoken the word or words corresponding to the hypothesis (step 1235). The recognizer makes this determination by comparing the current score for the hypothesis to a threshold value. If the score exceeds the threshold value, then the recognizer 215 determines that the hypothesis is too unlikely to merit further consideration and deletes the hypothesis (step 1240).
  • the recognizer determines whether the last word of the hypothesis is ending (step 1245). The recognizer determines that a word is ending when the frame corresponds to the last component of the model for the word. If the recognizer determines that a word is ending (step 1245), the recognizer sets a flag that indicates that the next frame may correspond to the beginning of a word (step 1250).
  • the recognizer selects the next hypothesis (step 1260) and repeats the comparison (step 1225) and other steps. If there are no more hypotheses to be considered for the frame (step 1255), then the recognizer determines whether there are more frames to be considered for the utterance (step 1265). The recognizer determines that there are more frames to be considered when two conditions are met. First, more frames must be available.
  • the best-scoring node for the current frame or for one or more of a predetermined number of immediately preceding frames must have been a node other than the silence node (that is, the utterance has ended when the silence node is the best-scoring node for the current frame and for a predetermined number of consecutive preceding frames).
  • step 1275 If there are more frames to be considered (step 1265) and the flag indicating that a word has ended is set (step 1270), or if there were no hypotheses to be considered for the frame (step 1215), then the recognizer requests from the pre-filtering procedure 240 a list of words that may start with the next frame (step 1275).
  • the recognizer then adds the modified list score to the score for the hypothesis and compares the result to a threshold value. If the result is less than the threshold value, then the recognizer maintains the hypothesis. Otherwise, the recognizer determines that the hypothesis does not merit further consideration and abandons the hypothesis. As an additional part of creating or expanding the hypotheses, the recognizer compares the hypotheses to the active constraint grammars 225 and abandons any hypothesis that corresponds to no active constraint grammar. The recognizer then retrieves the next frame (step 1210) and repeats the procedure.
  • the recognizer 215 provides the most likely hypotheses to the control/interface module 220 as recognition candidates (step 1285).
  • the control/interface module 220 controls operation of the speech recognition software and provides an interface to other software or to the user.
  • the control/interface module receives the list of recognition candidates for each utterance from the recognizer. Recognition candidates may correspond to dictated text, speech recognition commands, or external commands.
  • the control/interface module provides the text to an active application, such as a word processor.
  • the control/interface module also may display the best-scoring recognition candidate to the user through a graphical user interface.
  • the control/interface module 220 implements the command. For example, the control/interface module may control operation of the speech recognition software in response to speech recognition commands (for example, "wake up", "make that"), and may forward external commands to the appropriate software.
  • the control/interface module also controls the active vocabulary, acoustic models, and constraint grammars that are used by the recognizer. For example, when the speech recognition software is being used in conjunction with a particular application (for example, Microsoft Word), the control/interface module updates the active vocabulary to include command words associated with that application and activates constraint grammars associated with the application.
  • a particular application for example, Microsoft Word
  • control/interface module 220 Other functions provided by the control/interface module 220 include an enrollment program, a vocabulary customizer, and a vocabulary manager.
  • the enrollment program collects acoustic information from a user and trains or adapts a user's models based on that information.
  • the vocabulary customizer optimizes the language model of a specific topic by scanning user supplied text.
  • the vocabulary manager is a developer tool that is used to browse and manipulate vocabularies, grammars and macros.
  • Each function of the control/interface module 220 may be implemented as an executable program that is separate from the main speech recognition software.
  • the enrollment program may operate in an interactive mode that guides the user through the enrollment process, or in a non-interactive mode that permits the user to enroll independently of the computer.
  • the enrollment program displays the enrollment text to the user and the user reads the displayed text.
  • the recognizer 215 uses the enrollment grammar to match a sequence of utterances by the user to sequential portions of the enrollment text.
  • the enrollment program prompts the user to repeat certain passages of the text.
  • the recognizer uses acoustic information from the user's utterances to train or adapt acoustic models 235 corresponding to the matched portions of the enrollment text.
  • the interactive enrollment program is discussed in U.S. Application No. 08/825,536, titled "ENROLLMENT IN SPEECH RECOGNITION" and filed March 28, 1997.
  • the control/interface module 220 also implements error correction and cursor/position manipulation procedures of the software 124.
  • Error correction procedures include a "make that” command and a “spell that” command.
  • Cursor/position manipulation procedures include the "select” command discussed above and variations thereof (for example, “select [start] through [end]"), "insert before/after” commands, and a "resume with” command.
  • the various correction procedures are described in U.S. Patent No. 5,794,189, issued August 11, 1998 and U.S. Patent No. 6,064,959, issued May 16, 2000.
  • a complete dictation vocabulary 1300 includes the active vocabulary 230 plus a backup dictionary 245.
  • the backup dictionary 245 may include user-specific backup vocabulary words and system-wide backup vocabulary words.
  • User-specific backup vocabulary words include words that a user has added while using the speech recognition software. In one implementation, there is a limit of 64,000 user-specific words.
  • User-specific words are stored in vocabulary files for the user and for the dictation topic used during their addition, and are available as part of the backup dictionary for the dictation topic regardless of the user, and to the user regardless of the dictation topic.
  • any other user of the medical topic will have immediate access to the word "ganglion".
  • the word will be written into the user-specific backup vocabulary, so that, if the user later says “ganglion” while using a legal topic, the word "ganglion'' will be available.
  • the system-wide backup vocabulary words include all other non-active words known to the system.
  • the control/interface module 220 implements error correction procedures of the speech recognition software 124.
  • the speech recognition system 100 makes a recognition error
  • the user may invoke an appropriate correction command to remedy the error.
  • word searches of the backup dictionary 245 start with the user-specific backup dictionary and then check the system-wide backup dictionary.
  • the backup dictionary 245 also is searched when there are new words in text that a user has typed.
  • the backup dictionary 245 includes substantially more words than are included in the active vocabulary 230.
  • the backup dictionary 245 may have roughly 190,000 entries.
  • the active vocabulary 230 is a dynamic vocabulary in that entries may be added or subtracted from the active vocabulary over time. For example, when the user indicates that an error has been made and the control/interface module 220 uses the backup dictionary 245 to correct the error, a new word from the backup dictionary 245 may be added to the active vocabulary 230 to reduce the likelihood that the error will be repeated.
  • Word fragments include prefixes and suffixes, which are collectively known as affixes.
  • a prefix is any word fragment that precedes another word fragment or word and therefore may be represented as "prefix+", For example, a common prefix used in the English language is "re+”.
  • a suffix is any word fragment that follows another word fragment or word.
  • a suffix may be represented as ''+suffix". For example, a common suffix used in the English language is "+ed”.
  • Word fragments also include roots that can join with a prefix or a suffix.
  • a root may be a true root, in which case it is not a complete word when used alone and therefore may be represented as "+root+”.
  • An example of a true root is "+crimin+” which may form the word “decriminalize” when preceded by prefix "de+” and followed by the suffixes, in succession, "+al” and "+ize”.
  • a root also may be a word root that is a word when used alone. For example, the word root "look” may be used as the word "look” or may be joined with other word fragments to form another word such as "looked”.
  • Word fragments may also include separators, such as the "s" used in the German language to join words together or the hyphen, which is frequently used in the English language, and infixes, which are word fragments positioned in the interior of a word in such languages as Hungarian Mon-Khmer.
  • tags such as the "s" used in the German language to join words together or the hyphen, which is frequently used in the English language
  • infixes which are word fragments positioned in the interior of a word in such languages as Hungarian Mon-Khmer.
  • tags means “rest”
  • p -n- ak means "a support” such as a chair.
  • -n- could represent an infix in the Thai language.
  • Infixes are rare in the English language.
  • An example of an informal infix used in English might be -bloody- as in "fan -bloody- tastic".
  • Each word fragment 1315 may be included as a separate entry in the active vocabulary 230 along with a set of active words 1320 that typically includes a large portion of the previously active words 1312.
  • the word fragments 1315 increase the effective size of the active vocabulary to include words that may be in the backup dictionary 245, designated as splittable words 1325, that may be formed from combinations of word fragments with other word fragments or with active words.
  • the words of the backup dictionary that may not be formed from these combinations are designated as non-splittable words 1330.
  • the system When the recognizer 215 detects a word fragment 1315 in an utterance, the system combines the word fragment with one or more other word fragments or words to produce a new word, and then searches the backup dictionary 245 or applies rules to confirm that the new word is a valid word.
  • the speech recognition system 100 is able to tap into the backup dictionary 245 without requiring the recognizer to process words of the backup dictionary 245.
  • the recognition coverage of the speech recognition system 100 is expanded without significantly increasing the required memory capacity or processing time.
  • the backup dictionary 245 may include words that are not recognized as a concatenation of a sequence that includes words and word fragments. For example, many proper names may be included in the backup dictionary. The name "Scarlotti" is one example of a proper name that may not be splittable.
  • Figs. 14A and 15A show examples of the size of the complete dictation vocabulary 1300 of one implementation for, respectively, the French language and the U.S. version of the English language, without the use of word fragments.
  • the complete French dictation vocabulary 1400 includes about 180,000 words with about 40,000 of those words being in the French active vocabulary 230
  • the complete U.S. English dictation vocabulary 1500 includes about 240,000 words with about 30,000 of those words being in the U.S. English active vocabulary 230.
  • word fragments may be used to increase the effective size of the French active vocabulary 230, while keeping the complete dictation vocabulary 1400 near about 180,000 words and the active vocabulary 230 near 40,000 elements.
  • the active vocabulary 230 includes about 36,700 active French words 1402 and about 3,300 French word fragments 1405, such that the word fragments 1405 total about 8% of the total active vocabulary 230.
  • a substantial portion of the backup dictionary constitutes splittable words 1410 that may be accessed from the active vocabulary 230, while a much smaller portion constitutes non-splittable words 1415.
  • the French word fragments 1405 include mostly French true roots 1420 (that is, roots that are not words), while French word roots 1425 (that is, roots that are also words), French suffixes 1430, and French prefixes 1435 constitute a small segment of the French word fragments 1405.
  • word fragments also may be used to increase the effective size of the U.S. English active vocabulary 230, while keeping the complete dictation vocabulary 1500 near about 240,000 words and the active vocabulary 230 at 30,000 elements.
  • the active vocabulary 230 includes about 28,800 active U.S. English words 1502 and about 1,200 U.S. English word fragments 1505, such that the word fragments 1505 total about 4% of the total active vocabulary 230.
  • a smaller portion of the backup dictionary constitutes splittable words 1510 that may be accessed from the active vocabulary 230, while a larger portion constitutes non-splittable words 1515.
  • Fig. 15C the set of U.S.
  • U.S. English suffixes 1530 and prefixes 1535 constitute about one third of the U.S. English word fragments 1505, a substantially larger portion than the true roots.
  • the set of U.S. English word fragments 1505 may be larger or may include a larger percentage of true roots.
  • An effective size of the active vocabulary may be increased by creating an using fragmented word models. Memory and processing speed requirements tend to increase with the number of entries in the active vocabulary. As such, the size of the active vocabulary that may be processed in an allotted time by a particular processor using an available amount of memory is limited. Since the recognizer does not recognize words that are not included in the active vocabulary, the ability of the recognizer to recognize less-frequently-used words may be improved by increasing the size of the active vocabulary.
  • the effective size of the active vocabulary may by increased by dividing words into constituent components or fragments (for example, prefixes, suffixes, separators, infixes, and roots) and including each component as a separate entry in the active vocabulary.
  • a list of words and their plural forms for example, “book, books, cook, cooks, hook, hooks, look, and looks”
  • the words for example, "book, cook, hook, and look”
  • an entry representing the suffix that makes the words plural for example, "+s", where the "+” preceding the "s” indicates that "+s" is a suffix.
  • this technique may considerably reduce the number of vocabulary entries needed to represent the list of words.
  • word fragments may be generated in the speech recognition system 100 through a procedure 1600.
  • This procedure 1600 is implemented "offline"; that is before the speech recognition software 160 is sold and installed.
  • the procedure is performed by a system other than the speech recognition system 100.
  • word fragments 1315 are identified (step 1605). Then, using the word fragments 1315, a new active vocabulary 230 and a new backup dictionary 245 are generated (step 1610).
  • the new active vocabulary is constrained to be the same size as the old active vocabulary, such that the word fragments 1315 replace the least frequently used words of the active vocabulary.
  • the new French active vocabulary 230 has the same total number of elements (that is, 40,000) as the old French active vocabulary shown in Fig. 14A, but now includes word fragments 1405 and a corresponding reduction in the number of active words.
  • the new active vocabulary may be larger or smaller in size than the old active vocabulary.
  • the language model is created (step 1615). This permits the speech recognition system 100 to use the created language model to recognize both word fragments and words, and, in particular, to recognize word fragments in the context of surrounding word fragments and words.
  • the speech recognition system 100 is ready to perform speech recognition using the updated language model and the new active vocabulary according to an "on-line" procedure 1650. To this end, the system 100 performs speech recognition on an utterance to generate a list of candidates (step 1655).
  • the recognizer 215 operates according to the procedure 1200 shown in Fig. 12, with the exception that the recognizer 21 5 is able to detect word fragments 1315 as well as words.
  • the system 100 After the recognizer 215 generates the list of candidates for an utterance, the system 100 performs a post-recognition procedure (step 1660). During the post-recognition procedure, the system 100 combines any word fragments 1315 that may be included in a recognition candidate to form a word and looks up the word in the backup dictionary 245 or applies rules to verify that it is a valid word. Candidates that include word fragments that form invalid words are discarded.
  • the speech recognition system 100 performs speech recognition using the updated language model and the new active vocabulary (step 1655).
  • word fragments are treated like normal words, and the speech recognizer 215 operates according to the procedure 1200 illustrated in Fig. 12.
  • the recognizer 215 determines that there are no more speech frames to process, it provides the most likely hypotheses for the utterance to the control/interface module 220 as a list of recognition candidates (step 1285).
  • a particular recognition candidate may include word fragments in addition to words from the active vocabulary.
  • the speech recognition system 100 performs the post-recognition procedure 1660 that substitutes words from the backup dictionary or accepted by some set of rules for word fragments in a recognition candidate and rescores the recognition candidate in view of the substitution. Rescored candidates are placed on a revised candidate list. Candidates that do not include word fragments are transferred directly to the revised candidate list. Thus, the revised list may include both recognition candidates from the original list and rescored recognition candidates.
  • Homophones are words that sound the same but differ in spelling or representation, origin, and meaning.
  • many words are homophones.
  • French words include suffix word fragments that indicate the gender, number, person, and/or tense of the word.
  • the speech recognition systems described above fail to correctly distinguish homophones such as are used in the French language.
  • a word is a pure homophone of another word if the words are pronounced exactly the same. For example, the word “manger” is pronounced exactly like its pure homophones “mangé” and “mangés”.
  • a word is a near homophone of another word if some of the phonemes in the words sound very close. For example, the word “manger” sounds similar to its near homophone “mangeait”. As another example, the word “jouait” sounds similar to its near homophone “joué”.
  • a set of n words is an n-to-m homophone of a set of m words if the sets are pronounced the same.
  • the speech recognition system may process recognition results according to a procedure 1700.
  • the speech recognition system receives recognition results for an utterance (step 1705). While the recognition results may include multiple recognition candidates, the system processes only the best-scoring recognition candidate in implementing the procedure 1700. For ease of discussion, this candidate is referred to as the top recognition candidate.
  • the speech recognition system determines whether the top recognition candidate includes only text (step 1715). If so, the speech recognition system processes the text (step 1720). For example, the speech recognition system may insert the text 1800, 1805 into a dictation window 1810, as shown in Figs. 18A and 18B.
  • the speech recognition system determines whether the candidate includes a correction command, such as the "make that” or “spell that” commands discussed above (step 1725). If the top recognition candidate does not include a correction command (step 1725), the speech recognition system processes any other command included in the top recognition candidate (step 1730). For example, if the top recognition candidate includes text and a formatting command such as New-Paragraph, then the speech recognition system processes the text portion of the candidate and performs the formatting command.
  • a correction command such as the "make that” or "spell that” commands discussed above.
  • the speech recognition system determines whether the candidate includes a homophone correction command that permits active correction of homophone mistakes (step 1735).
  • the homophone correction command may be implemented as, for example, "Write ⁇ word> ⁇ ending>" in English or "Ecrire ⁇ word> ⁇ ending>” in French.
  • the ⁇ word> is a word in the visible text and the ⁇ ending> is a finite set of possible endings (for example, about 300 possible endings).
  • the user spells out the ending. In other implementations, the user may type or pronounce the ending.
  • step 1735 If the candidate does not include a homophone correction command (step 1735), then the speech recognition system performs a review of the best-scoring recognition candidates (step 1740) before displaying a correction dialog box (step 1745) and entering correction mode (step 1750).
  • the speech recognition system performs a procedure 1740 for reviewing the best-scoring recognition candidates. Initially, the speech recognition system selects the best-scoring recognition candidate (step 1900) and determines whether the best-scoring recognition candidate includes any homophones (step 1905). If the best-scoring recognition candidate does not include one or more homophones (step 1905), then the speech recognition system selects the next best-scoring recognition candidate (step 1910) to determine if the that best-scoring recognition candidate includes one or more homophones (step 1905).
  • the speech recognition system retrieves all possible grammatical categories (that is, parts of speech) using a tag set associated with the word (step 1915).
  • a tag set (written as ⁇ P 1 P 2 ... P n ⁇ ) for a word includes tags corresponding to P i , grammatical categories of the word.
  • Grammatical categories include, for example, parts of speech such as nouns, prepositions, adjectives, verbs, pronouns, and features of words, where features include, for example, person, gender, number, and tense.
  • P 3 may include information about person (for example, second person) and P 8 may include information about tense (for example, past tense).
  • the speech recognition system retrieves all possible grammatical categories 2000 for a selected recognition candidate "nous nous la premises moulds” 2005 using tags 2010.
  • the tag for the word "nous” is ⁇ _4PF_T_ ⁇ and the tag for the word "sommes” is ⁇ _V4NEAE_ ⁇ .
  • the possible grammatical categories 2000 for the word "nous” are OD (direct object pronoun), OI (indirect object pronoun), PerR (personal reflexive pronoun), PerS (personal subject pronoun), and PerT (personal pronoun), and the possible grammatical categories 2000 for the word "sommes” are NV (noun) and VC (conjugated verb).
  • the speech recognition system retrieves all possible grammatical categories 2100 for a selected recognition candidate "quelque cinquanteurs cutaneous tissues" 2105 using tags 2110.
  • the tag for the word "echoesente” is (_E8_BT_____) and the tag for the word "zzi” is ⁇ /V6_BE____ ⁇ .
  • the only possible grammatical category for the word "cinquante” is NumC (number) and the possible grammatical categories for the word "quelque” are Adv (adverb) and Adj (adjective).
  • the speech recognition system guesses a grammatical category for each word in the selected recognition candidate (1920).
  • the speech recognition system uses rules to designate a grammatical category for a word in the selected recognition candidate (this may be referred to as resolving grammatical conflicts). These rules include local syntactic or contextual information, as discussed below.
  • the word “le” in the selected recognition candidate “le bel yogurt jouait,” may be an object pronoun or a determinant.
  • the speech recognition system uses the rule that a determinant precedes an adjective (that is, "bel”) and guesses that the word "le” is a determinant.
  • the grammatical categories 2015 for each word in the selected recognition candidate are guessed using the rules 2017,
  • the grammatical categories 2115 for each word in the selected recognition candidate are guessed using the rules 2117.
  • one rule may be: a word that normally would be an adjective and that precedes a number must be an adverb.
  • This rule is written in Fig. 21A as " ⁇ Adv: "quelque”> (NumC, "peu")".
  • the word "quelque” in the selected recognition candidate 2105 precedes the word Neinquantew which is a number the word "quelque” must be an adverb.
  • the rule written in Fig. 20A as "nous” ⁇ Adv ⁇ ⁇ PerR: "nous”> indicates that if the word “nous” precedes the word "nous,” then the second instance of "nous” is grammatically a personal reflexive pronoun.
  • the speech recognition system retrieves a feature set for each word in the selected recognition candidate (step 1925).
  • the feature set for a word is the set of all relevant features, such as person, gender, number, and tense, for that word and may be considered a subset of the tag set.
  • the feature sets 2020 and 2120, respectively are retrieved for each of the words in the selected recognition candidates 2005 and 2105, respectively.
  • the speech recognition system arranges the words in the selected recognition candidate into phrases (step 1930).
  • a phrase is a combination of words that are related to each other. For example, in the selected recognition candidate "Ie bel Georgia jouait,” a noun phrase comprises “le bel yogurt” and a verb phrase comprises "jouait.”
  • the speech recognition system arranges the words into phrases based on the grammatical categories of each of the words.
  • Figs. 20B and 21B show word phrases 2025 and 2125 for the selected recognition candidates 2005 and 2105.
  • "cortors” is a verb phrase
  • "cinquante dealt” is a noun phrase
  • "la” is a noun phrase.
  • the speech recognition system determines the relationships among (or between) those phrases using a set of phrase rules (step 1935). For example, in the selected recognition candidate "le belieri jouait,” the speech recognition system uses the rule that a noun phrase is the subject of a verb phrase to determine that "le belura” is the subject of "jouait.”
  • Figs. 20B and 21B show phrase relationships 2030 and 2130 for the selected recognition candidates 2005 and 2105.
  • the word “nous” is the subject of the verb phrase "sommes boness.”
  • the noun phrase “cinquanteurs” is the subject of verb phrase "licas.”
  • the speech recognition system resolves conflicts in the feature sets (step 1940) by replacing each of the words in the selected recognition candidate by that word's homophones or near-homophones. For example, the speech recognition system may resolve the conflict in the selected recognition candidate "le belieri jouait” between the word “ieris” (which is plural) and the words “jouait,” “le,” and “bel,” which are singular, by replacing the word “ieris” with its singular homophone "enfant.”
  • the speech recognition system resolves the conflict between the word “partagés,” which is masculine in gender and plural in number, and the word “la,” which is feminine in gender and singular in number, by replacing the word “partagés” with the word “partagee,” which is feminine in gender and singular in number.
  • Fig. 20B shows the words 2035 in the candidate after conflicts are resolved.
  • the speech recognition system resolves the conflict between the words “personnes” and “cinquante,” which are plural in number and the word “echoesente,” which is singular in number, by replacing the word “echoesente” with the word “geniesentes,” which is plural in number. In this way, all of the relevant words in the selected recognition candidate 2105 agree in number. Also shown in Fig. 21B are the words 2135 in the candidate after conflicts are resolved.
  • the speech recognition system replaces a word in the selected recognition candidate with homophones or near-homophones of that word if the replacement word fixes the conflict in the feature sets.
  • the speech recognition system determines which homophones in the selected recognition candidate should be replaced using the feature set information of non-homophone words in the selected recognition candidate. In this way, the speech recognition system may rely on results that were most likely correct. For example, in the selected recognition candidate "nous nous la premises answerss," the words “nous,” “la,” and “sommes" are non-homophones.
  • the speech recognition system uses the rule that a past participle must inherit the feature set of its object when replacing the word “partagés” with the word “partagée” because "la” is an object of the past participle "partagés.” Because “la” is feminine and singular, “partagés” must also be feminine and singular. Therefore, the speech recognition system replaces "partagés" with "partagée.
  • the speech recognition system When the conflict is resolved, the speech recognition system outputs the alternative recognition candidate as the selected recognition candidate (step 1945). In one implementation, the speech recognition system assigns a better score to the alternative recognition candidate than the selected recognition candidate because the user presumably wanted to correct the best-scoring recognition candidate.
  • the speech recognition system determines if the choice list is full (step 1950). If not, the speech recognition system determines if there are any other recognition candidates from the choice list that contain homophones (step 1955). If there are other recognition candidates that contain homophones (step 1955), the speech recognition system selects the next best-scoring recognition candidate for review (step 1910). If the choice list if full (step 1950) or if there are no other recognition candidates that contain homophones (step 1955), then, referring again to Fig. 17, the speech recognition system presents the alternative choice list (made up of original recognition candidates and/or alternative recognition candidates) to the user (step 1745) by, for example, displaying the alternative choice list in a correction dialog box.
  • the speech recognition system performs the procedure 1740 in a way that is imperceptible to the user.
  • the speech recognition system may not always shift the feature sets of homophones in a phrase correctly. This may happen because there may be several interpretations of a homophone. This may also happen if there are no non-homophones in the selected recognition candidate. When this happens, the speech recognition system would be unable to compare the feature set of a homophone to a non-homophone. Thus, the speech recognition system would align all feature sets of the homophones (for example, replace homophones having mixed feature sets with homophones having a single feature set). In this case, the speech recognition system may perform an incorrect alignment.
  • the speech recognition system begins processing of the homophone correction command by finding a previously-recognized utterance that includes the ⁇ word> from the homophone correction command (step 1770).
  • the command is limited to require the ⁇ word> to be one that is visible on the display. This limitation simplifies the task of finding the appropriate utterance to be corrected.
  • the utterance to be corrected is modified by replacing ⁇ word> with a variation of ⁇ word> that includes ⁇ ending> (step 1775).
  • the speech recognition system then processes the modified utterance (step 1780) by resolving conflicts in feature sets of phrases in the best-scoring recognition candidate as discussed above with reference to Fig. 19.
  • the features of the homophone formed from the ⁇ word> and ⁇ ending> should agree with features of the relevant non-homophone words in the best-scoring recognition candidate. If the features are not in agreement, then the speech recognition system modifies the recognition candidate in accordance with the ⁇ word> and ⁇ ending> such that the features of the relevant non-homophone words agree with the features of the replacement homophone.
  • the utterance to be corrected 1800 is "ce thank les amis que j' ai aidés et protégés" and the user has spoken the homophone correction command "ecrire amis ies” 1820.
  • the speech recognition system finds the root “amis'' in the utterance to be corrected 1800 and determines the word formed by the root "amis" and the ending "ies," which, in this example is "amies.”
  • the speech recognition system may review all of the words in the candidate 1800 to determine if there are any other homophones that conflict in features with “amies.”
  • the speech recognition system determines that replacing "amis,” which is masculine, with “amies,” which is feminine, requires the masculine words “aidés” and “protégés” to be replaced with the feminine homophones “alibers” and “protégées,” respectively, to resolve conflicts in the gender feature.
  • the utterance to be corrected 1805 is "par laloid de cours .s fermés qui irrigina à" and the user has entered the homophone correction command 1825 "ecrire attribuée ées.”
  • the speech recognition system combines the root “attribuée” with the ending "ées” to form the word “attribuées.”
  • the speech recognition system initially processed "lalus” as the subject of "insulating attribué.” However, because the user corrected the word "attribué,” the speech recognition system resolves the conflict and processes "cours tes fermées” as the subject of " standing Samuel attribueés.”
  • the speech recognition system determines (step 1785) that there are multiple possible alternative homophone recognition results, then the speech recognition system displays the correction dialog box (step 1745). For example, the speech recognition system may determine that more than one word may be formed from the word "attribuée” and the word "ées,” in which case there may be several different alternative homophone recognition candidates with resolved conflicts.
  • the speech recognition system After display of the correction dialog box (step 1745), the speech recognition system enters correction mode (step 1750) and processes the best-scoring recognition candidate and input from the user (step 1755) until the system determines that correction is complete (step 1760). If correction is complete (step 1760), the speech recognition system exits correction mode and clears the correction dialog box (step 1765). The system then waits for further recognition results (step 1705).
  • step 1785 determines (step 1785) that there are not multiple alternative homophone recognition results, and if the speech recognition system determines there is only one possible alternative homophone recognition result based on processing (step 1790), then the speech recognition system replaces the utterance to be corrected with the alternative homophone recognition result (step 1792) and waits for receipt of the next set of recognition candidates (step 1705).
  • the speech recognition system identifies no alternative homophone recognition results, this indicates that the combination of ⁇ word> and ⁇ ending> does not generate a homophone. In this case, the speech recognition system notifies the user (step 1795) that the homophone correction command was unsuccessful. Optionally, the speech recognition system may display a correction dialog box.
  • the techniques described here are not limited to any particular hardware or software configuration; they may find applicability in any computing or processing environment that may be used for speech recognition.
  • the techniques may be implemented in hardware or software, or a combination of the two.
  • the techniques are implemented in computer programs executing on programmable computers that each include a processor, a storage medium readable by the processor (including volatile and non-volatile memory and/or storage elements), at least one input device, and at least one output device.
  • Program code is applied to data entered using the input device to perform the functions described and to generate output information.
  • the output information is applied to one or more output devices.
  • Each program is preferably implemented in a high level procedural or object oriented programming language to communicate with a computer system.
  • the programs can be implemented in assembly or machine language, if desired.
  • the language may be a compiled or interpreted language.
  • Each such computer program is preferably stored on a storage medium or device (for example, CD-ROM, hard disk or magnetic diskette) that is readable by a general or special purpose programmable computer for configuring and operating the computer when the storage medium or device is read by the computer to perform the procedures described in this document.
  • a storage medium or device for example, CD-ROM, hard disk or magnetic diskette
  • the system may also be considered to be implemented as a computer-readable storage medium, configured with a computer program, where the storage medium so configured causes a computer to operate in a specific and predefined manner.
  • the speech recognition system only generates one alternative homophone choice for each choice provided.
  • the alternative is given a better score so that it will be preferred to the original choice and appear during dictation.
  • the old choice is not destroyed or penalized.
  • the alternative is neither acoustically rescored nor rescored using language models.
  • the speech recognition system may generate more than a single alternative recognition candidate, and the original recognition candidate may not be penalized.
  • the speech recognition system may also implement the procedure 1700 with n-to-m homophones.
  • the homophone correction command is "write ⁇ word> ⁇ ending>.”
  • the homophone correction command may be "write ⁇ homophone> ⁇ affix>” or "write ⁇ affix> ⁇ homophone>,” where ⁇ homophone> represents n words and ⁇ affix> is the prefix or suffix that should be combined with the homophone to form the corrected homophone.

Landscapes

  • Engineering & Computer Science (AREA)
  • Computational Linguistics (AREA)
  • Health & Medical Sciences (AREA)
  • Audiology, Speech & Language Pathology (AREA)
  • Human Computer Interaction (AREA)
  • Physics & Mathematics (AREA)
  • Acoustics & Sound (AREA)
  • Multimedia (AREA)
  • Machine Translation (AREA)
  • Character Discrimination (AREA)
  • Telephonic Communication Services (AREA)
  • Telephone Function (AREA)

Claims (30)

  1. Verfahren zum Korrigieren eines fehlerhaften Textes in Verbindung mit Erkennungsfehlern in der computerunterstützten Spracherkennung, wobei das Verfahren umfasst:
    Durchführen einer Spracherkennung zu einer Äußerung, um ein Erkennungsergebnis für die Äußerung zu erzeugen;
    Identifizieren eines Korrekturbefehls in dem Erkennungsergebnis für die Äußerung, wobei der Korrekturbefehl umfasst:
    eine Gruppe aus ein oder mehreren Worten eines früheren Erkennungsergebnisses, und
    ein Affix, das mit der Gruppe aus ein oder mehreren Worten kombiniert werden kann,
    um ein Homophon der ein oder mehreren Worte zu bilden; und
    Erzeugen des Homophons aus der Gruppe von ein oder mehreren Worten und des Affix basierend auf den Merkmalen ein oder mehrerer Wörter des früheren Erkennungsergebnisses, die nicht in der Gruppe aus ein oder mehreren Worten enthalten sind.
  2. Verfahren nach Anspruch 1, mit einem Ersetzen der Gruppe aus ein oder mehreren Worten des früheren Erkennungsergebnisses durch das Homophon.
  3. Verfahren nach Anspruch 1, in welchem das Erzeugen des Homophons aus der Gruppe aus ein oder mehreren Wörtern und des Affix ein Kombinieren der Gruppe von ein oder mehreren Worten und des Affix in der Weise umfasst, dass die Merkmale der ein oder mehreren Wörter des früheren Erkennungsergebnisses mit dem Homophon zusammenpassen.
  4. Verfahren nach Anspruch 1, in welchem das Erzeugen des Homophons umfasst:
    Präsentieren einer Liste möglicher Homophone; und
    Erhalten einer Auswahl eines Homophons aus der Liste.
  5. Verfahren nach Anspruch 1, in welchem der Korrekturbefehl umfasst "ecrire <wort><endung>", wobei <wort> die Gruppe von ein oder mehrere Worten des früheren Erkennungsergebnisses repräsentiert und <endung> das Affix repräsentiert.
  6. Verfahren nach Anspruch 5, in weichem die Gruppe aus ein oder mehreren Worten nur ein Wort umfasst und das Affix eine Endung ist.
  7. Verfahren nach Anspruch 1, in welchem das Homophon genauso wie die Gruppe aus ein oder mehreren Worten des früheren Erkennungsergebnisses betont wird.
  8. Verfahren nach Anspruch 1, in welchem ein oder mehrere Phoneme des Homophons in der Gruppe aus ein oder mehreren Worten des früheren Erkennungsergebnisses enthalten sind.
  9. Verfahren nach Anspruch 1, in welchem das Homophon eine Gruppe aus ein oder mehreren Worten umfasst.
  10. Verfahren nach Anspruch 1, in welchem das Affix eine Endung ist.
  11. Computerunterstütztes Spracherkennungssystem mit:
    einem Prozessor;
    einer Eingabeeinrichtung, die eine Äußerung erhält;
    einer Speichereinrichtung zum Speichern von Softwarebefehlen, die von dem Prozessor durchgeführt werden;
    Mitteln zum Durchführen einer Spracherkennung zu der Äußerung, um ein Erkennungsergebnis für die Äußerung zu erzeugen;
    Mitteln zum identifizieren eines Korrekturbefehls in dem Erkennungsergebnis für die Äußerung, wobei der Korrekturbefehl umfasst:
    eine Gruppe aus ein oder mehreren Worten eines früheren Erkennungsergebnisses, und
    ein Affix, das mit der Gruppe aus ein oder mehreren Worten kombiniert werden kann, um ein Homophon der ein oder mehreren Worte zu bilden; und
    Mittel zum Erzeugen des Homophons aus der Gruppe aus ein oder mehreren Worten und des Affix basierend auf Merkmalen von ein oder mehreren Worten des früheren
    Erkennungsergebnisses, die nicht in der Gruppe aus ein oder mehreren Worten enthalten sind.
  12. System nach Anspruch 11, in welchem die Speichereinrichtung von dem Prozessor ausgeführte Softwarebefehle zum Ersetzen der Gruppe aus ein oder mehreren Worten des früheren Erkennungsergebnisses durch das Homophon speichert.
  13. System nach Anspruch 11, in welchem das Mittel zum Erzeugen des Homophons aus der Gruppe aus ein oder mehreren Worten und des Affix Mittel zum Kombinieren der Gruppe aus ein oder mehreren Worten und des Affix umfasst, derart, dass die Merkmale der ein oder mehreren Worte des früheren Erkennungsergebnisses mit dem Homophon zusammenpassen.
  14. System nach Anspruch 11, in welchem die Mittel zum Erzeugen des Homophons Mittel umfassen:
    zum Präsentieren einer Liste möglicher Homophone; und
    zum Erhalten einer Auswahl eines Homophons aus der Liste.
  15. System nach Anspruch 11, in welchem der Korrekturbefehl umfasst "ecrire <wort><endung>", wobei <wort> die Gruppe aus ein oder mehreren Worten des früheren Erkennungsergebnisses repräsentier und <endung> das Affix repräsentiert.
  16. System nach Anspruch 15, in welchem die Gruppe aus ein mehreren Worten nur ein Wort enthält und das Affix eine Endung ist.
  17. System nach Anspruch 11, in welchem das Homophon genauso wie die Gruppe aus ein oder mehreren Worten des früheren Erkennungsergebnisses betont wird.
  18. System nach Anspruch 11, in welchem ein oder mehrere Phoneme des Homophons in der Gruppe aus ein oder mehreren Worten des früheren Erkennungsergebnisses enthalten sind.
  19. System nach Anspruch 11, in welchem das Homophon eine Gruppe aus ein oder mehreren Worten enthält.
  20. System nach Anspruch 11, in welchem das Affix eine Endung ist.
  21. Computersoftware, merklich ausgeführt in einem computerlesbaren Medium oder in einem mitlaufenden Trägersignal, zum Korrigieren eines fehlerhaften Textes in Verbindung mit Erkennungsfehlern, wobei die Software die folgenden Codesegmente umfasst, um ihre entsprechenden Operationen durchzuführen, wenn sie auf einem Computersystem läuft:
    ein erstes Codesegment, um eine Spracherkennung zu der Äußerung durchzuführen, um ein Erkennungsergebnis für die Äußerung zu erzeugen;
    ein zweites Codesegment, um einen Korrekturbefehl in dem Erkennungsergebnis für die Äußerung zu identifizieren, wobei der Korrekturbefehl umfasst:
    eine Gruppe aus ein oder mehreren Worten eines früheren Erkennungsergebnisses, und
    ein Affix, das mit der Gruppe aus ein oder mehreren Worten kombiniert werden kann,
    um ein Homophon der ein oder mehreren Worte zu bilden; und
    ein drittes Codesegment, um das Homophon aus der Gruppe aus ein oder mehreren Worten und des Affix basierend auf Merkmalen der ein oder mehreren Worte des früheren Erkennungsergebnisses zu erzeugen, die nicht in der Gruppe aus ein oder mehreren Worten enthalten sind.
  22. Software nach Anspruch 21, ferner mit einem vierten Codesegment, um die Gruppe aus ein oder mehreren Worten des früheren Erkennungsergebnisses durch das Homophon zu ersetzen.
  23. Software nach Anspruch 21, in welcher das dritte Codesegment ein Codesegment umfasst, um die Gruppe aus ein oder mehreren Worten und das Affix derart zu kombinieren, dass die Merkmale der ein oder mehreren Worte des früheren Erkennungsergebnisses mit dem Homophon zusammenpassen.
  24. Software nach Anspruch 21, in welcher das dritte Codesegment umfasst:
    ein Codesegment, um eine Liste möglicher Homophone zu präsentieren; und
    ein Codesegment, um eine Auswahl eines Homophons aus der Liste zu erhalten.
  25. Software nach Anspruch 21, in welcher der Korrekturbefehl umfasst: "ecrire <wort><endung>", wobei <wort> die Gruppe aus oder mehreren Worten des früheren Erkennungsergebnisses repräsentiert und <endung> das Affix repräsentiert.
  26. Software nach Anspruch 25, in welcher die Gruppe aus ein oder mehreren Worten nur ein Wort umfasst und das Affix eine Endung ist.
  27. Software nach Anspruch 21, in welcher das Homophon genauso wie die Gruppe aus ein oder mehreren Worten des früheren Erkennungsergebnisses betont wird.
  28. Software nach Anspruch 21, in welcher ein oder mehrere Phoneme des Homophons in der Gruppe aus ein oder mehreren Worten des früheren Erkennungsergebnisses enthalten sind.
  29. Software nach Anspruch 21, in welcher das Homophon eine Gruppe aus ein oder mehreren Worten enthält.
  30. Software nach Anspruch 21, in welcher das Affix eine Endung ist.
EP01307935A 2000-09-18 2001-09-18 Homophonewahl in der Spracherkennung Expired - Lifetime EP1189203B1 (de)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
US66454500A 2000-09-18 2000-09-18
US664545 2000-09-18

Publications (3)

Publication Number Publication Date
EP1189203A2 EP1189203A2 (de) 2002-03-20
EP1189203A3 EP1189203A3 (de) 2002-11-20
EP1189203B1 true EP1189203B1 (de) 2006-05-17

Family

ID=24666410

Family Applications (1)

Application Number Title Priority Date Filing Date
EP01307935A Expired - Lifetime EP1189203B1 (de) 2000-09-18 2001-09-18 Homophonewahl in der Spracherkennung

Country Status (3)

Country Link
EP (1) EP1189203B1 (de)
AT (1) ATE326754T1 (de)
DE (1) DE60119643T2 (de)

Families Citing this family (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
DE10251112A1 (de) * 2002-11-02 2004-05-19 Philips Intellectual Property & Standards Gmbh Verfahren und System zur Spracherkennung
GB2409561A (en) * 2003-12-23 2005-06-29 Canon Kk A method of correcting errors in a speech recognition system
US20060293889A1 (en) * 2005-06-27 2006-12-28 Nokia Corporation Error correction for speech recognition systems
CN103076893B (zh) * 2012-12-31 2016-08-17 百度在线网络技术（北京）有限公司 一种用于实现语音输入的方法与设备
US9514743B2 (en) 2014-08-29 2016-12-06 Google Inc. Query rewrite corrections
CN113223516B (zh) * 2021-04-12 2022-11-29 北京百度网讯科技有限公司 语音识别方法和装置

Family Cites Families (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
GB9511855D0 (en) * 1995-06-12 1995-08-09 Dragon Syst Uk Ltd Speech recognition apparatus and methods
US5752230A (en) * 1996-08-20 1998-05-12 Ncr Corporation Method and apparatus for identifying names with a speech recognition program
CA2303312A1 (en) * 1997-09-24 1999-04-01 Guido Gallopyn Apparatus and method for distinguishing similar-sounding utterances in speech recognition

Also Published As

Publication number Publication date
DE60119643T2 (de) 2007-02-01
DE60119643D1 (de) 2006-06-22
ATE326754T1 (de) 2006-06-15
EP1189203A3 (de) 2002-11-20
EP1189203A2 (de) 2002-03-20

Similar Documents

Publication Publication Date Title
EP1346343B1 (de) Spracherkennung durch wort-in-phrase-befehl
EP0867857B1 (de) Registrierung für die Spracherkennung
US6064959A (en) Error correction in speech recognition
US6424943B1 (en) Non-interactive enrollment in speech recognition
US6601027B1 (en) Position manipulation in speech recognition
US7120582B1 (en) Expanding an effective vocabulary of a speech recognition system
US6092044A (en) Pronunciation generation in speech recognition
US7584102B2 (en) Language model for use in speech recognition
EP0867859B1 (de) Sprachmodelle für die Spracherkennung
US7315818B2 (en) Error correction in speech recognition
US7580838B2 (en) Automatic insertion of non-verbalized punctuation
EP0376501B1 (de) Spracherkennungssystem
Geutner et al. Adaptive vocabularies for transcribing multilingual broadcast news
US6345249B1 (en) Automatic analysis of a speech dictated document
EP1189203B1 (de) Homophonewahl in der Spracherkennung
EP1204966A1 (de) Erweiterung eines effektiven wortschatzes eines spracherkennungsystems

Legal Events

Date Code Title Description
PUAI Public reference made under article 153(3) epc to a published international application that has entered the european phase

Free format text: ORIGINAL CODE: 0009012

AK Designated contracting states

Kind code of ref document: A2

Designated state(s): AT BE CH CY DE DK ES FI FR GB GR IE IT LI LU MC NL PT SE TR

AX Request for extension of the european patent

Free format text: AL;LT;LV;MK;RO;SI

PUAL Search report despatched

Free format text: ORIGINAL CODE: 0009013

AK Designated contracting states

Kind code of ref document: A3

Designated state(s): AT BE CH CY DE DK ES FI FR GB GR IE IT LI LU MC NL PT SE TR

AX Request for extension of the european patent

Free format text: AL;LT;LV;MK;RO;SI

17P Request for examination filed

Effective date: 20030515

AKX Designation fees paid

Designated state(s): AT BE CH CY DE DK ES FI FR GB GR IE IT LI LU MC NL PT SE TR

GRAP Despatch of communication of intention to grant a patent

Free format text: ORIGINAL CODE: EPIDOSNIGR1

GRAS Grant fee paid

Free format text: ORIGINAL CODE: EPIDOSNIGR3

GRAA (expected) grant

Free format text: ORIGINAL CODE: 0009210

AK Designated contracting states

Kind code of ref document: B1

Designated state(s): AT BE CH CY DE DK ES FI FR GB GR IE IT LI LU MC NL PT SE TR

PG25 Lapsed in a contracting state [announced via postgrant information from national office to epo]

Ref country code: IT

Free format text: LAPSE BECAUSE OF FAILURE TO SUBMIT A TRANSLATION OF THE DESCRIPTION OR TO PAY THE FEE WITHIN THE PRESCRIBED TIME-LIMIT;WARNING: LAPSES OF ITALIAN PATENTS WITH EFFECTIVE DATE BEFORE 2007 MAY HAVE OCCURRED AT ANY TIME BEFORE 2007. THE CORRECT EFFECTIVE DATE MAY BE DIFFERENT FROM THE ONE RECORDED.

Effective date: 20060517

Ref country code: CH

Free format text: LAPSE BECAUSE OF FAILURE TO SUBMIT A TRANSLATION OF THE DESCRIPTION OR TO PAY THE FEE WITHIN THE PRESCRIBED TIME-LIMIT

Effective date: 20060517

Ref country code: NL

Free format text: LAPSE BECAUSE OF FAILURE TO SUBMIT A TRANSLATION OF THE DESCRIPTION OR TO PAY THE FEE WITHIN THE PRESCRIBED TIME-LIMIT

Effective date: 20060517

Ref country code: LI

Free format text: LAPSE BECAUSE OF FAILURE TO SUBMIT A TRANSLATION OF THE DESCRIPTION OR TO PAY THE FEE WITHIN THE PRESCRIBED TIME-LIMIT

Effective date: 20060517

Ref country code: AT

Free format text: LAPSE BECAUSE OF FAILURE TO SUBMIT A TRANSLATION OF THE DESCRIPTION OR TO PAY THE FEE WITHIN THE PRESCRIBED TIME-LIMIT

Effective date: 20060517

Ref country code: BE

Free format text: LAPSE BECAUSE OF FAILURE TO SUBMIT A TRANSLATION OF THE DESCRIPTION OR TO PAY THE FEE WITHIN THE PRESCRIBED TIME-LIMIT

Effective date: 20060517

Ref country code: FI

Free format text: LAPSE BECAUSE OF FAILURE TO SUBMIT A TRANSLATION OF THE DESCRIPTION OR TO PAY THE FEE WITHIN THE PRESCRIBED TIME-LIMIT

Effective date: 20060517

REG Reference to a national code

Ref country code: GB

Ref legal event code: FG4D

REG Reference to a national code

Ref country code: CH

Ref legal event code: EP

REG Reference to a national code

Ref country code: IE

Ref legal event code: FG4D

REF Corresponds to:

Ref document number: 60119643

Country of ref document: DE

Date of ref document: 20060622

Kind code of ref document: P

PG25 Lapsed in a contracting state [announced via postgrant information from national office to epo]

Ref country code: SE

Free format text: LAPSE BECAUSE OF FAILURE TO SUBMIT A TRANSLATION OF THE DESCRIPTION OR TO PAY THE FEE WITHIN THE PRESCRIBED TIME-LIMIT

Effective date: 20060817

Ref country code: DK

Free format text: LAPSE BECAUSE OF FAILURE TO SUBMIT A TRANSLATION OF THE DESCRIPTION OR TO PAY THE FEE WITHIN THE PRESCRIBED TIME-LIMIT

Effective date: 20060817

PG25 Lapsed in a contracting state [announced via postgrant information from national office to epo]

Ref country code: ES

Free format text: LAPSE BECAUSE OF FAILURE TO SUBMIT A TRANSLATION OF THE DESCRIPTION OR TO PAY THE FEE WITHIN THE PRESCRIBED TIME-LIMIT

Effective date: 20060828

PG25 Lapsed in a contracting state [announced via postgrant information from national office to epo]

Ref country code: IE

Free format text: LAPSE BECAUSE OF NON-PAYMENT OF DUE FEES

Effective date: 20060918

PG25 Lapsed in a contracting state [announced via postgrant information from national office to epo]

Ref country code: MC

Free format text: LAPSE BECAUSE OF NON-PAYMENT OF DUE FEES

Effective date: 20060930

PG25 Lapsed in a contracting state [announced via postgrant information from national office to epo]

Ref country code: PT

Free format text: LAPSE BECAUSE OF FAILURE TO SUBMIT A TRANSLATION OF THE DESCRIPTION OR TO PAY THE FEE WITHIN THE PRESCRIBED TIME-LIMIT

Effective date: 20061017

ET Fr: translation filed
NLV1 Nl: lapsed or annulled due to failure to fulfill the requirements of art. 29p and 29m of the patents act
REG Reference to a national code

Ref country code: CH

Ref legal event code: PL

PLBE No opposition filed within time limit

Free format text: ORIGINAL CODE: 0009261

STAA Information on the status of an ep patent application or granted ep patent

Free format text: STATUS: NO OPPOSITION FILED WITHIN TIME LIMIT

26N No opposition filed

Effective date: 20070220

GBPC Gb: european patent ceased through non-payment of renewal fee

Effective date: 20060918

PG25 Lapsed in a contracting state [announced via postgrant information from national office to epo]

Ref country code: GB

Free format text: LAPSE BECAUSE OF NON-PAYMENT OF DUE FEES

Effective date: 20060918

PG25 Lapsed in a contracting state [announced via postgrant information from national office to epo]

Ref country code: GR

Free format text: LAPSE BECAUSE OF FAILURE TO SUBMIT A TRANSLATION OF THE DESCRIPTION OR TO PAY THE FEE WITHIN THE PRESCRIBED TIME-LIMIT

Effective date: 20060818

PG25 Lapsed in a contracting state [announced via postgrant information from national office to epo]

Ref country code: TR

Free format text: LAPSE BECAUSE OF FAILURE TO SUBMIT A TRANSLATION OF THE DESCRIPTION OR TO PAY THE FEE WITHIN THE PRESCRIBED TIME-LIMIT

Effective date: 20060517

Ref country code: LU

Free format text: LAPSE BECAUSE OF NON-PAYMENT OF DUE FEES

Effective date: 20060918

PG25 Lapsed in a contracting state [announced via postgrant information from national office to epo]

Ref country code: CY

Free format text: LAPSE BECAUSE OF FAILURE TO SUBMIT A TRANSLATION OF THE DESCRIPTION OR TO PAY THE FEE WITHIN THE PRESCRIBED TIME-LIMIT

Effective date: 20060517

REG Reference to a national code

Ref country code: FR

Ref legal event code: PLFP

Year of fee payment: 16

REG Reference to a national code

Ref country code: FR

Ref legal event code: PLFP

Year of fee payment: 17

PGFP Annual fee paid to national office [announced via postgrant information from national office to epo]

Ref country code: FR

Payment date: 20171229

Year of fee payment: 17

PG25 Lapsed in a contracting state [announced via postgrant information from national office to epo]

Ref country code: FR

Free format text: LAPSE BECAUSE OF NON-PAYMENT OF DUE FEES

Effective date: 20180930

PGFP Annual fee paid to national office [announced via postgrant information from national office to epo]

Ref country code: DE

Payment date: 20191129

Year of fee payment: 19

REG Reference to a national code

Ref country code: DE

Ref legal event code: R119

Ref document number: 60119643

Country of ref document: DE

PG25 Lapsed in a contracting state [announced via postgrant information from national office to epo]

Ref country code: DE

Free format text: LAPSE BECAUSE OF NON-PAYMENT OF DUE FEES

Effective date: 20210401